Research Article
Amharic Character Recognition Using Deep Convolutional Neural Network
@INPROCEEDINGS{10.1007/978-3-031-28725-1_11, author={Achamie Aynalem}, title={Amharic Character Recognition Using Deep Convolutional Neural Network}, proceedings={Artificial Intelligence and Digitalization for Sustainable Development. 10th EAI International Conference, ICAST 2022, Bahir Dar, Ethiopia, November 4-6, 2022, Proceedings}, proceedings_a={ICAST}, year={2023}, month={3}, keywords={Amharic recognition Character recognition Deep convolutional neural networks Deep learning application}, doi={10.1007/978-3-031-28725-1_11} }- Achamie Aynalem
Year: 2023
Amharic Character Recognition Using Deep Convolutional Neural Network
ICAST
Springer
DOI: 10.1007/978-3-031-28725-1_11
Abstract
Amharic is the working language in the Federal Democratic Republic of Ethiopia. The Amharic alphabet has a large number of symbols and there is a close resemblance among shapes of the different symbols available in the language which challenged the task of machine-based optical character recognition systems in the language. The absence of a standardized labeled dataset for the Amharic language created additional barriers for different researchers. Our aim in this paper is to design a deep convolutional neural network based architecture that could extract features and classify Amharic characters with significant confidence of accuracy that could be utilized for real-world applications. A total of 90,000 characters are prepared for training the proposed architecture and an additional of 25,000 characters are reserved for testing purpose. Due to the occurrence of a large number of symbols and a close resemblance in the shapes of the different characters available in the language, a relatively complex convolutional Neural Network is utilized to capture those features and categorize them into the correct characters. Dropout layers are utilized to avoid overfitting. The character recognition system proposed in this paper achieved an accuracy of 99.27% on the testing dataset which is a significant improvement for the Amharic language. The implementation was done using Tensorflow on Keras neural network layers and Opencv in python to pre-process image data which enables us to make the system readily available for software developers as an API.
